Gate the model on speech instead of transcribing everything

[z33b0t]

Fixes background chatter being transcribed while the HUD showed no voice activity.

The bug

The VAD's verdict decided when to cut a segment and what the HUD displayed, but never whether to decode. Every frame was buffered and handed to Parakeet, which has no VAD of its own — so a room's background conversation came back as words while the HUD sat dimmed.

Three paths fed the model unconditionally, and gating only the first would have been worse than useless:

path	effect of leaving it ungated
live partial	noise appears in the live view
segment commit	noise lands in the committed transcript
one-shot final re-decode of the whole dictation	noise comes back when you press stop

So the gate goes where the audio is buffered, and all three inherit it.

The fix

SileroVAD.process now returns the audio worth keeping; the session buffers that instead of the raw frame. Padding keeps the gate from eating real speech — Silero crosses its threshold a beat after a word starts and falls below it while the last consonant is still sounding:

• VAD_PREROLL_MS (250) of held audio flushed ahead of each onset, as a ring of the most recent rejected audio so an hour of silence still costs 250 ms
• VAD_HANGOVER_MS (200) kept past each offset
• VAD_THRESHOLD (0.5) is now settable — it was a constructor default no call site passed, so tuning the knob most likely to need tuning per-room meant editing vad.py and rebuilding the image

SpeechGate holds the padding rules with no model and no torch import, so CI can test it without a GPU. Also adds SileroVAD.flush(): whole-window classification always holds back <32 ms, which is exactly the end of your last word when you stop mid-sentence.

Verification

New gate suite: 5 unit checks on the padding rules plus an end-to-end dictation of nothing but rejected audio asserting neither partials nor the final carry text. FakeVAD now drives the real SpeechGate, so what CI covers is production code.

Measured against the real Silero + Parakeet in the container, driving it with espeak speech over the WebSocket. Sweeping amplitude, checking whether text ever came out of audio the VAD never flagged:

amplitude	peak	vad fired	text
1.000	25685	yes	full sentence
0.020	513	yes	full sentence
0.004	102	yes	first word lost
0.002	51	yes	(empty)
0.001	25	no	(empty)

No row pairs "didn't fire" with text. At normal dictation volume the VAD fires 68 ms after speech starts and the pre-roll opens the gate 156 ms before it; at −48 dBFS the VAD lags 436 ms, which is what VAD_PREROLL_MS is for.

Trade-off worth knowing: the gate costs accuracy on very quiet speech. At those bottom two rows the ungated /v1 path transcribes both in full. Those levels are far below anything you'd dictate at, but it isn't free.

Running live on the dictation server for 32 hours at the 0.5 default, confirmed by the reporter as fixing the original symptom.

Side effects

• full is now speech-only, so FINAL_MAX_S bounds speaking time rather than wall-clock — a long, mostly-quiet dictation that used to exceed it now gets the better one-shot final.
• /v1 uploads get the threshold but are never gated: handing us a file is itself a decision that its contents are worth transcribing. Makes _segment_bounds' "same thresholds the live path uses" true rather than aspirational.

🤖 Generated with Claude Code